Optimizing the prediction of local soil fertility parameters using Brazilian national and regional spectral libraries comprising XRF and vis-NIR data

Abstract

This project will evaluate and advance predictive modeling strategies associate with soil spectral libraries involving sensor data from X-ray fluorescence spectroscopy (XRF) and diffuse reflectance spectroscopy in the visible and near infrared regions (vis-NIR) for the diagnosis of tropical soil fertility. Essentially, three questions that comprise the state of the art for spectral libraries of XRF and vis-NIR sensors will be answered, they are (i) what is the best strategy to calibrate predictive models using the abovementioned spectral libraries to determine soil fertility of local samples? (ii) which tropical soil attributes can be predicted with satisfactory performance when using these spectral libraries?; (iii) does the association of the spectral library with non-linear models using Random Forest (RF) allow better predictions than linear partial least squares regression (PLSR) models? These evaluations will be performed in a national spectral library composed of 2605 samples from 54 different Brazilian agricultural fields and in a regional spectral library (for the State of São Paulo) comprising 997 samples from 16 agricultural areas. We have chosen one of the most outstanding groups in the field of pedometrics and data science applied to spectral libraries to carry out this project in partnership. The project will be conducted in three sequential steps, in which the questions i and ii will be answered in steps 1 and 2; step 1 evaluating the national spectral library and step 2 evaluating the regional spectral library. Finally, question iii will be answered in step 3. (AU)